Internet access performance in LTE TDD
Riikka Susitaival1, Henning Wiemann3, Jessica stergaard2, Anna Larmo1
Ericsson Research
1
Finland, 2Sweden, 3Germany
[riikka.susitaival, henning.wiemann, jessica.ostergaard, anna.larmo]@ericsson.com
Abstract The Time Division Duplex (TDD) uplink-downlink
configuration of the 3GPP Long Term Evolution (LTE)
determines how the ten subframes in a radio frame are divided
between the downlink and the uplink. The specified
configurations cover a wide range of allocations from a
downlink-heavy resource distribution ratio (9:1) to an uplink-
heavy ratio (2:3). In this paper, we compare the performance
of Internet access using the TCP protocol in different
downlink-heavy asymmetries. We find that the performance
depends on many factors such as the transferred file size, the
control channel errors and the downlink/uplink traffic mix.
When the file size is small, TDD can not fully utilize its
potentially higher downlink capacity because of longer uplink
access delays as well as shortage of uplink resources in the
chosen configurations. With an increased file size, this effect
fades away and TDD provides higher download bit rates than
FDD. The realized increase in bit rate is however not as high as
the calculated increase in available downlink resources.
Keywords- LTE, TDD, Internet access, TCP
I. INTRODUCTION
The first release (Release 8) of the Long Term Evolution
(LTE) has recently been standardized by 3GPP. LTE
provides high peak data rates up to 300 Mbps, improved
spectrum efficiency, and reduced radio access delays. One
key requirement in the development of LTE has been
spectrum flexibility; LTE can be operated in different
spectrum allocations from 1.4 to 20 MHz and in paired or
unpaired spectrum [1]. The Frequency Division Duplex
(FDD) mode uses paired spectrum where different carrier
frequencies are used for downlink (DL) and uplink (UL).
The Time Division Duplex (TDD) mode, on the other hand,
uses unpaired spectrum and a single carrier frequency with
separation of downlink and uplink in time. In TDD, a
subframe in a radio frame can either be an uplink, a
downlink or a special subframe. 3GPP has specified seven
different uplink-downlink configurations.
The possibility to distribute TDD radio resources
unevenly among uplink and downlink has many
consequences. Depending on traffic characteristics and
asymmetries, it is possible to dimension uplink/downlink
capacity in such a way that the available spectrum is utilized
most efficiently. This is done by selecting an appropriate
TDD configuration. In order to avoid interference between
uplink and downlink, such as base station to base station and
mobile to mobile interference, the network is preferably
synchronized and all macro cells within a geographical area
use the same TDD configuration. Hence, the configuration
should be set considering the average load in the system.
978-1-4244-2519-8/10/$26.00 2010 IEEE
Another characteristic of LTE TDD is that when the number
of downlink subframes is different from the number of
uplink subframes, there is no direct one-to-one mapping
between uplink and downlink subframes. This has an effect
on the Hybrid Automatic Repeat reQuest (HARQ) operation
of LTE TDD as well as on the L1/L2 control signaling
making it different compared to FDD.
This work concentrates on Internet access using the
Transmission Control Protocol (TCP) [7]. In the case of a
download, the data goes in the downlink and control data
(TCP acknowledgements (ACKs)) in the uplink. The initial
TCP congestion window is small and thus the radio link is
typically not fully utilized in this first phase of the data
transfer. The TCP performance suffers from high delays in
this phase as they prevent the congestion window from
increasing quickly, i.e., the link from being further utilized.
Once the congestion window of the TCP transmitter exceeds
the bandwidth-delay product of the link, TCP is rather
insensitive to delays.
In this paper, we study downlink-heavy TDD
configurations which aim at providing improved downlink
performance. First, the HARQ aspects of LTE TDD are
introduced and user level latencies are evaluated. Using
advanced system simulations, we then study the actual
performance of TDD and compare it to FDD. We use both
pure download as well as mixed downlink/uplink traffic
models. In addition, we study how L1/L2 control signaling
errors affect the performance in TDD.
So far, specific features and lower-layer performance
aspects have been studied and published for LTE TDD. For
example, Astly et al. simulate TDD configuration 1 in [1],
and Rahman and Astly study TDD ACK-NACK bundling
with link level simulations in [2]. In contrast to previous
studies, we consider also higher layer protocols aspects and
two-directional traffic models using both uplink and
downlink in this paper. TCP performance studies focusing on
TDD asymmetries exist for WiMAX ([3], [4]), but due to
inherent differences in the radio access network
architectures, as well as TDD frame structures, these results
cannot directly be applied to an LTE network.
This paper is organized as follows: in Section II, the
TDD frame structure is described briefly. HARQ related
issues such as user plane latency and ACK-NACK bundling,
are evaluated in Sections III and IV. In Section V, the
simulation model is presented and Section VI studies TDD
performance by simulations. Finally, conclusions are drawn
in Section VII.
 II. TDD FRAME STRUCTURE
Similar to FDD, the TDD radio frame consists of 10
subframes, each having a length of 1 ms. In TDD, the
subframe can be either DL, UL, or a special subframe
between the DL and UL period. 3GPP has defined seven
uplink-downlink configurations. The DL:UL patterns of the
configurations are listed in Table I. For example, in
configuration 1, there are always three consecutive DL
subframes (including both the normal DL subframes and the
special subframes), followed by two UL subframes. The
same pattern is then repeated starting from DL. See Figure 1
for illustration of configuration 1 in more detail.
TABLE I. DOWNLINK-UPLINK CONFIGURATIONS SPECIFIED BY 3GPP.
Configuration DL:UL
0 2:3
1 3:2
2 4:1
3 7:3
4 8:2
5 9:1
6 3:3:2:2
UL Special subframe
DL
#0 #1 #2 #3 #4 #5 #6 #7 #8 #9
Figure 1. TDD uplink downlink Configuration 1, TDD DL:UL 3:2.
The special subframe consists of a downlink part
(DwPTS), a guard period (GP), and an uplink part (UpPTS).
The DL part of the special subframe can be considered as a
normal downlink subframe for data and control but with a
reduced number of data symbols. The guard period consists
of a number of idle symbols where nothing is transmitted to
protect against DL-UL interference. Finally, the UL part of
the special subframe is considerably shorter than the
downlink part and is only used for channel sounding and
random access preamble transmission (no user data
transmission). 3GPP has defined nine different con-
figurations for how the 14 symbols of the special subframe
are divided between the downlink part, the uplink part, and
the guard period. The special subframe configurations are
listed in the 3GPP specification [6].
III. HARQ TIMING AND USER PLANE LATENCIES
IN TDD MODE
The multiprocess stop-and-wait HARQ mechanism used
in LTE to enable a low-error-rate data stream is based on a
known delay between transmission of data and transmission
of feedback. The feedback can be either a positive
acknowledgement (ACK) or a negative acknowledgement
(NACK). In FDD, the delay between the data and the
feedback transmission is always 4 ms. In TDD, if there is a
DL transmission in subframe n, the subframe n+4 is not
necessarily an UL subframe. Thus, the feedback has to be
delayed. The same problem appears for uplink traffic and the
corresponding downlink HARQ feedback. The exact HARQ
timing relations for each TDD configuration have therefore
been specified by 3GPP and can be found in [5].
Latencies play an important role in end-to-end
performance of user applications. In the following, the
transmission latency in uplink is studied and compared
between FDD and TDD. It is assumed that the User Equip-
ment (UE) is time-aligned, and that it has been configured
with dedicated resources for sending a Scheduling Request
(SR) to ask for an uplink transmission grant, but that it does
not yet have the grant for transmission.
The procedure to transmit data, e.g., TCP ACK, in the
uplink after data has arrived to the UEs buffer is as follows:
1) In the subframe where the UE has an SR resource
available, the UE transmits an SR, which is a one-bit flag to
indicate that the UE has new data. 2) The eNB receives the
SR and after a processing delay, an initial grant is
transmitted to the UE allocating frequency resources for UL
transmission. The processing delay of the SR is selected to 3
ms which is similar to HARQ processing times specified for
the UE. 3) Using the granted resources, the UE transmits
data as well as a Buffer Status Report (BSR) to indicate to
the eNodeB how much data it still has available in its buffer
after the transmission. The delay from the grant transmission
by the eNodeB to the uplink transmission by the UE is fixed
to 4 ms in FDD but in TDD, it varies between uplink-
downlink configurations. 4) When the eNB has received the
BSR, it can continue allocating uplink resources to the UE
and the UE can perform further uplink transmissions.
Figure 2 depicts the uplink data transmission procedure
for FDD and the TDD 3:2 configuration. From the diagrams
we can see that the delay from the SR transmission to the
second scheduled transmission from which the UE can be
scheduled continuously based on the buffer information, is
16 ms in FDD whereas in the studied TDD configuration, it
is 20 ms. It can be calculated that this delay is 20 ms also for
most of the other TDD configurations.
IV. ACK-NACK BUNDLING
In addition to HARQ timing and user plane latencies,
there is another major difference between TDD and FDD:
with downlink-heavy configurations, the UE must transmit
multiple HARQ feedbacks in one uplink subframe. 3GPP has
standardized two mechanisms to achieve that: ACK-NACK
bundling and ACK-NACK multiplexing. In ACK-NACK
multiplexing, all HARQ feedbacks are explicitly signaled in
one uplink subframe and since more bits are transmitted, the
required SINR is higher. In this study, ACK/NACK bundling
is used and thus introduced in more detail.
 SR Initial grant TX1 Grant TX2

#0 #2 #4 #6 #8 #10 #12 #14 #16 #18

SR Initial grant TX1 Grant TX2

#0 #2 #4 #6 #8 #10 #12 #14 #16 #18 #20 #22

Figure 2. The user plane latencies in FDD (upper figure) and TDD 3:2 (lower figure). Blue subrames are for DL and orange for UL.

In ACK-NACK bundling, a logical AND is done over the transport protocols. The path loss model is a typical distance-
HARQ feedback messages. If all DL HARQ processes in a so- dependent model with wrap around.
called bundling window are correctly decoded, an ACK is sent
by the UE. If one or more of the processes are not correctly The downlink-heavy TDD configurations, 3:2, 7:3, 4:1 and
decoded, the UE transmits NACK and all processes in the 9:1, are studied. The 8:2 configuration is excluded because it
window have to be retransmitted. In this context, the bundling offers very similar DL:UL ratio to TDD 4:1. We choose
window refers to those DL subframes whose feedback configuration 7 for the special subframe including 12 DL
messages are reported in a single UL subframe. If a OFDM symbols. The main parameters are listed in Table II.
multicodeword transmission is used for any of the subframes in
a bundling window, then a bundled feedback is sent per TABLE II. SIMULATION ASSUMPTIONS.
codeword. ACK-NACK bundling with the TDD 3:2
configuration is illustrated in Figure 3. The feedback timing Parameter Value
relations for all TDD allocations are defined in [5]. System Bandwidth TDD: 10 MHz, FDD: 5 + 5 MHz

TX
Feed- Feed- Network layout Hexagonal grid, 21 cells
TX 2 TX 3 back1 back2
Site-to-site distance 500 m

Multipath model Typical Urban, 3 km/h

Antenna model SIMO, 1 transmit and 2 receive

#0 #2 #4 #6 #8 antennas in both uplink and downlink
Max. DL TX power FDD: 20 W TDD: 40 W
AND
Max. UL TX power FDD and TDD: 0.25 W
Figure 3. ACK/NACK bundling with TDD 3:2. The first bundling window is
over subframes #0 and #1 and the second over subframe #4. Scheduling Round robin
SR on uplink control 10 ms periodicity
In the case when multiple DL transmissions are scheduled
channel
by the eNB in a bundling window, the UE might miss one of
the assignments. The design of ACK-NACK bundling has HARQ transmissions DL: max 7, UL: max 8
considered this: The DL assignment carries a Downlink RLC mode Acknowledged (AM)
Assignment Index (DAI) indicating how many assignments the
UE should have received so far within the current bundling
window. If the UE detects that the DAI differs from the
number of correctly received DL assignments, it does not send
any HARQ feedback and the eNB can detect this. However, the VI. SIMULATION RESULTS
eNB cannot know which of the transmissions was missed, and In this section, results from multiple simulations are
thus the whole bundle has to be retransmitted. evaluated to compare the download performance of TDD and
FDD. As a theoretical background, in Table III the number of
V. SIMULATION MODEL DL and UL resource elements (RE) for data in a radio frame is
presented for FDD and TDD. The numbers exclude overhead
The simulations are done with a detailed system-level for reference signals and uplink and downlink control channels.
simulator implementing the radio link with 3GPP L1/L2 and
TABLE III. THE NUMBER OF RESOURCES ELEMENTS IN A RADIO FRAME. whereas the number of DL resource elements increased 78%
when comparing TDD 9:1 and FDD. This is due to many
DL REs UL REs
factors such as ACK/NACK bundling resulting in unnecessary
FDD 31500 27720 retransmissions, and UL limitations and delays.
TDD 3:2 36600 (+16 %) 22176 (-20 %) File size 1 MB
100
TDD 7:3 43500 (+38 %) 16632 (-40 %)
FDD
TDD 4:1 49200 (+56 %) 11088 (-60 %) 80 TDD 3:2
TDD 9:1 56100 (+ 78 %) 5544 (-80 %) TDD 7:3
60 TDD 4:1

CDF
A. Single user download peak rates TDD 9:1
40
First, download peak rates are studied. A single user at a
time is placed in the system for requesting and downloading a
20
single file with TCP. Once done, the user leaves the system and
a new user arrives. The file size varies from 100 kB to 10 MB.
0
The arrival position is selected randomly in the system 9 10 11 12 13
according to a uniform distribution. Object bit rate (Mbps)

Figure 4 shows the CDF for the download performance in Figure 5. The download object bit rate, file size 1 MB.
terms of the object bit rate, that is, the file size divided by the
download time, when the file size is small, 100 kB. Even
though the downlink capacity is larger in TDD, most of TDD File size 10 MB
configurations do not improve the performance when 100
compared to FDD. This can be explained by longer uplink FDD
access delays of TDD as studied in Section III. The TCP 80
TDD 3:2
congestion window increases exponentially for each received
60 TDD 7:3
TCP ACK. If these feedback messages are delayed due to
CDF

TDD 4:1
longer UL delays, the congestion window increases more
40 TDD 9:1
slowly and the object bit rate remains lower. In addition, from
Figure 4 it can be seen that TDD 9:1 has the worst performance
at the lower 10th percentile. This can be explained by too small 20
instantaneous uplink capacity and power, especially when the
UE is in bad radio conditions. TCP ACKs can arrive to the 0
0 5 10 15 20 25 30
uplink buffer in bursts and it takes a relatively long time to Object bit rate (Mbps)
empty the queue.
Figure 6. The download object bit rate, file size 10 MB.
File size 100 KB
100
B. Download mean object bit rates
FDD
80
TDD 3:2
In the following, we investigate the download performance in a
TDD 7:3
multi-user scenario. The users arrive to the system according to
60
TDD 4:1
a Poisson process with varying arrival intensity between
CDF

TDD 9:1
40
20
0 when there are more users in the system, each individual user
2.2 2.4 2.6 2.8 3 3.2
Object bit rate (Mbps)
Figure 4. The download object bit rate, file size 100 KByte.
With a larger file size, the TCP initial phase does not have
as much impact on the overall transfer time. Thus, uplink
limitations do not have such a big effect. When the file size is
medium, 1MB, in Figure 5, the object bit rate of TDD 4:1 is
highest but TDD 9:1 suffers still from uplink limitations.
Finally, with a large file size, 10 MB, in Figure 6, the extreme
configuration TDD 9:1 provides the highest bit rate. However,
the difference is not as large as the numbers in Table III may
indicate. The peak rate increases 62% in simulation results
scenarios. The file size is 1 MB. The user object bit rate as a
function of the DL throughput is depicted in Figure 7. As the
load and the number of users in the system increases (when
increasing the user arrival rate), the overall throughput, i.e., the
number of served bits, increases as well. On the other hand,
gets fewer resources and the object bit rate decreases. From
Figure 7, we can see that the higher the load of the system, the
bigger the difference between TDD and FDD. The increased
DL capacity can be better utilized when there are more users.
C. Impact of control channel signaling errors on download
In this subsection, the impact of L1/L2 control channel errors
on TDD download performance is studied. We focus on DL
HARQ related control channels. The errors are independent
and occur with probability p in each subframe. The values for p
are selected according to the 3GPP targets:
 uplink-heavy configuration, TDD will naturally handle higher
12
UL load.
Mean user object bit rate (Mbps)

FDD
10
TDD-3-2 12
8 TDD-7-3
FDD

Mean user object bit rate (Mbps)

TDD-4-1
10 TDD-3-2
TDD-9-1
6 TDD-7-3
TDD-4-1
8
4 TDD-9-1

2 6

0
0 1 2 3 4 5 6 4
Mean cell throughput (Mbps/cell)

Figure 7. The mean object bit rate as a function of cell throughput in multi-
user scenario for file download. File size 1 MByte.
1) DL assignment loss on Physical Downlink Control Channel (PDCCH)
(p=10-2). If this error occurs, the UE does not transmit any HARQ feedback.
The eNB does not receive any feedback, i.e., it detects DTX, and retransmits all
active DL HARQ processes of the corresponding ACK/NACK bundling
window.
2) DTX-to-ACK error on Physical Uplink Control Channel (PUCCH) (p=10 ).
This error leads to data loss of the bundling window at the HARQ level and an
RLC retransmission if operated in RLC acknowledged mode.
3) NACK-to-ACK error on PUCCH (p=10-4). This error also leads to data loss
at HARQ level and possible RLC retransmission.
In Table IV, the download object bit rates (median) in a single-
user scenario are depicted with and without the above control
channel errors. It can be seen that the impact of the errors is
higher with TDD as compared to FDD. This difference is due
to ACK-NACK bundling. One missed DL assignment leads to
the retransmission of all HARQ processes in the bundling
window. By operating the system such that the missed
assignment probability is lower, the losses due to missed
assignments can be made lower. Alternatively, the loss seen in
TDD can be controlled by configuration selection.
TABLE IV. 50% PERCENTILE DOWNLOAD OBJECT BIT RATE. SINGLE
USER SCENARIO, FILE SIZE 10 MB
Mode No errors With errors
FDD 15.5 Mbps 15.4 Mbps
TDD 3:2 17.3 Mbps 16.8 Mbps
TDD 7:3 20.5 Mbps 20.1 Mbps
TDD 4:1 22.5 Mbps 21.5 Mbps
TDD 9:1 24.5 Mbps 22.6 Mbps
D. Mixed traffic scenario
Finally, a mixed traffic scenario including both download
and upload users is studied. The file size is 1 MB and the
control channels are error free. The overall user arrival
intensity is fixed to 5 users per second but the fraction of
download and upload users is varying. In Figure 8, the user
object bit rate averaged over all users (download and upload) is
depicted as a function of the fraction of upload users. We can
see that when the fraction of upload users is over 40%, FDD
outperforms TDD. This is due to shortage of UL resources in
the chosen downlink-heavy TDD configurations. With an
2
0 10 20 30 40 50 60
Share of upload users (%)
Figure 8. The mean user object bit rate as a function of the share of
uploading users. File size 1 MB. The total user arrival rate is fixed.
VII. CONCLUSIONS
-2 In this paper, we studied the performance of Internet access
with TDD downlink-heavy asymmetries. By simulations, we
show that the performance depends on many factors such as the
file size and errors on the control channels. If the file size is
small, around 100 kB, longer user plane delays as well as
shortage of UL capacity and power does not allow getting the
full advantage of the asymmetry of TDD. With a larger file
size, the download bit rates can be improved as compared to
FDD, even though the improvement is not as significant as the
increase in resource allocation would suggest. We have also
studied a mixed traffic model with both downloading and
uploading users. In the studied scenario, FDD performs better
than TDD in terms of the mean user bit rate if the share of
uploading users is over 40%.
As discussed in the introduction, due to possible
interference between uplink and downlink, the TDD
configuration is preferably same over the geographical area.
Thus the adaptation to instantaneous traffic changes in the cell
Difference
is difficult and the large range of TDD configuration cannot
always be fully utilized. Instead, it is preferable to select the
-0.65 % configuration based on the long term traffic averages.
-2.9 % REFERENCES
-1.9 % [1] D. Astly, E. Dahlman, A. Furuskr, Y. Jading, M. Lindstrm, and S.
Parkvall, LTE: The Evolution of Mobile Broadband, IEEE
-4.4 % Communication Magazine, April 2009.
-7.8 % [2] M. Rahman, D. Astly, Link level investigation of ACK/NACK
bundling for LTE TDD, IEEE VTC Spring, 2009.
[3] A. Eshete, A. Arcia, D. Ros, and Y. Jiang, Impact of WiMAX Network
Assymetry on TCP, . IEEE WCNC, 2009.
[4] C-H Chiang, w. Liao, T. Liu, I. Chan, H-L Chao, Adaptive downlink
and uplink channel split ratio determination for TCP based best effort
traffic in TDD-based WiMAX networks, IEEE Journal on Selected
Areas in Communications, February 2009.
[5] 3GPP TS 36.213, Evolved Universal Terrestrial Radio Access (E-
UTRA); Physical Layer Procedures.
[6] 3GPP TS 36.211, Evolved Universal Terrestrial Radio Access (E-
UTRA); Physical channels and modulation.
[7] W. Stewens, TCP/IP illustrated, Volume 1, Addison-Wesley, 1994.